优化azure的体验

config.py

@@ -8,7 +8,7 @@
 """
 
 # [step 1]>> API_KEY = "sk-123456789xxxxxxxxxxxxxxxxxxxxxxxxxxxxxx123456789"。极少数情况下，还需要填写组织（格式如org-123456789abcdefghijklmno的），请向下翻，找 API_ORG 设置项
-API_KEY = "sk-此处填API密钥"    # 可同时填写多个API-KEY，用英文逗号分割，例如API_KEY = "sk-openaikey1,sk-openaikey2,fkxxxx-api2dkey1,fkxxxx-api2dkey2"
+API_KEY = "此处填API密钥"    # 可同时填写多个API-KEY，用英文逗号分割，例如API_KEY = "sk-openaikey1,sk-openaikey2,fkxxxx-api2dkey3,azure-apikey4"
 
 
 # [step 2]>> 改为True应用代理，如果直接在海外服务器部署，此处不修改
@@ -110,9 +110,8 @@ SLACK_CLAUDE_USER_TOKEN = ''
 
 # 如果需要使用AZURE 详情请见额外文档 docs\use_azure.md
 AZURE_ENDPOINT = "https://你亲手写的api名称.openai.azure.com/"
-AZURE_API_KEY = "填入azure openai api的密钥"
+AZURE_API_KEY = "填入azure openai api的密钥"    # 建议直接在API_KEY处填写，该选项即将被弃用
-AZURE_API_VERSION = "2023-05-15"        # 一般不修改
+AZURE_ENGINE = "填入你亲手写的部署名"            # 读 docs\use_azure.md
-AZURE_ENGINE = "填入你亲手写的部署名"     # 读 docs\use_azure.md
 
 
 # 使用Newbing

main.py

@@ -4,10 +4,10 @@ def main():
     import gradio as gr
     if gr.__version__ not in ['3.28.3','3.32.2']: assert False, "需要特殊依赖，请务必用 pip install -r requirements.txt 指令安装依赖，详情信息见requirements.txt"
     from request_llm.bridge_all import predict
-    from toolbox import format_io, find_free_port, on_file_uploaded, on_report_generated, get_conf, ArgsGeneralWrapper, DummyWith
+    from toolbox import format_io, find_free_port, on_file_uploaded, on_report_generated, get_conf, ArgsGeneralWrapper, load_chat_cookies, DummyWith
     # 建议您复制一个config_private.py放自己的秘密, 如API和代理网址, 避免不小心传github被别人看到
-    proxies, WEB_PORT, LLM_MODEL, CONCURRENT_COUNT, AUTHENTICATION, CHATBOT_HEIGHT, LAYOUT, API_KEY, AVAIL_LLM_MODELS, AUTO_CLEAR_TXT = \
+    proxies, WEB_PORT, LLM_MODEL, CONCURRENT_COUNT, AUTHENTICATION, CHATBOT_HEIGHT, LAYOUT, AVAIL_LLM_MODELS, AUTO_CLEAR_TXT = \
-        get_conf('proxies', 'WEB_PORT', 'LLM_MODEL', 'CONCURRENT_COUNT', 'AUTHENTICATION', 'CHATBOT_HEIGHT', 'LAYOUT', 'API_KEY', 'AVAIL_LLM_MODELS', 'AUTO_CLEAR_TXT')
+        get_conf('proxies', 'WEB_PORT', 'LLM_MODEL', 'CONCURRENT_COUNT', 'AUTHENTICATION', 'CHATBOT_HEIGHT', 'LAYOUT', 'AVAIL_LLM_MODELS', 'AUTO_CLEAR_TXT')
 
     # 如果WEB_PORT是-1, 则随机选取WEB端口
     PORT = find_free_port() if WEB_PORT <= 0 else WEB_PORT
@@ -54,7 +54,7 @@ def main():
     cancel_handles = []
     with gr.Blocks(title="ChatGPT 学术优化", theme=set_theme, analytics_enabled=False, css=advanced_css) as demo:
         gr.HTML(title_html)
-        cookies = gr.State({'api_key': API_KEY, 'llm_model': LLM_MODEL})
+        cookies = gr.State(load_chat_cookies())
         with gr_L1():
             with gr_L2(scale=2):
                 chatbot = gr.Chatbot(label=f"当前模型：{LLM_MODEL}")

request_llm/bridge_all.py

@@ -16,9 +16,6 @@ from toolbox import get_conf, trimmed_format_exc
 from .bridge_chatgpt import predict_no_ui_long_connection as chatgpt_noui
 from .bridge_chatgpt import predict as chatgpt_ui
 
-from .bridge_azure_test import predict_no_ui_long_connection as azure_noui
-from .bridge_azure_test import predict as azure_ui
-
 from .bridge_chatglm import predict_no_ui_long_connection as chatglm_noui
 from .bridge_chatglm import predict as chatglm_ui
 
@@ -48,10 +45,11 @@ class LazyloadTiktoken(object):
         return encoder.decode(*args, **kwargs)
 
 # Endpoint 重定向
-API_URL_REDIRECT, = get_conf("API_URL_REDIRECT")
+API_URL_REDIRECT, AZURE_ENDPOINT, AZURE_ENGINE = get_conf("API_URL_REDIRECT", "AZURE_ENDPOINT", "AZURE_ENGINE")
 openai_endpoint = "https://api.openai.com/v1/chat/completions"
 api2d_endpoint = "https://openai.api2d.net/v1/chat/completions"
 newbing_endpoint = "wss://sydney.bing.com/sydney/ChatHub"
+azure_endpoint = AZURE_ENDPOINT + f'openai/deployments/{AZURE_ENGINE}/chat/completions?api-version=2023-05-15'
 # 兼容旧版的配置
 try:
     API_URL, = get_conf("API_URL")
@@ -122,9 +120,9 @@ model_info = {
 
     # azure openai
     "azure-gpt-3.5":{
-        "fn_with_ui": azure_ui,
+        "fn_with_ui": chatgpt_ui,
-        "fn_without_ui": azure_noui,
+        "fn_without_ui": chatgpt_noui,
-        "endpoint": get_conf("AZURE_ENDPOINT"),
+        "endpoint": azure_endpoint,
         "max_token": 4096,
         "tokenizer": tokenizer_gpt35,
         "token_cnt": get_token_num_gpt35,

request_llm/bridge_azure_test.py

@@ -1,237 +0,0 @@
-"""
-    该文件中主要包含三个函数
-
-    不具备多线程能力的函数：
-    1. predict: 正常对话时使用，具备完备的交互功能，不可多线程
-
-    具备多线程调用能力的函数
-    2. predict_no_ui：高级实验性功能模块调用，不会实时显示在界面上，参数简单，可以多线程并行，方便实现复杂的功能逻辑
-    3. predict_no_ui_long_connection：在实验过程中发现调用predict_no_ui处理长文档时，和openai的连接容易断掉，这个函数用stream的方式解决这个问题，同样支持多线程
-"""
-
-import logging
-import traceback
-import importlib
-import openai
-import time
-import requests
-import json
-
-# 读取config.py文件中关于AZURE OPENAI API的信息
-from toolbox import get_conf, update_ui, clip_history, trimmed_format_exc
-TIMEOUT_SECONDS, MAX_RETRY, AZURE_ENGINE, AZURE_ENDPOINT, AZURE_API_VERSION, AZURE_API_KEY = \
-    get_conf('TIMEOUT_SECONDS', 'MAX_RETRY',"AZURE_ENGINE","AZURE_ENDPOINT", "AZURE_API_VERSION", "AZURE_API_KEY")
-
-
-def get_full_error(chunk, stream_response):
-    """
-        获取完整的从Openai返回的报错
-    """
-    while True:
-        try:
-            chunk += next(stream_response)
-        except:
-            break
-    return chunk
-
-def predict(inputs, llm_kwargs, plugin_kwargs, chatbot, history=[], system_prompt='', stream = True, additional_fn=None):
-    """
-    发送至azure openai api，流式获取输出。
-    用于基础的对话功能。
-    inputs 是本次问询的输入
-    top_p, temperature是chatGPT的内部调优参数
-    history 是之前的对话列表（注意无论是inputs还是history，内容太长了都会触发token数量溢出的错误）
-    chatbot 为WebUI中显示的对话列表，修改它，然后yeild出去，可以直接修改对话界面内容
-    additional_fn代表点击的哪个按钮，按钮见functional.py
-    """
-
-    if additional_fn is not None:
-        import core_functional
-        importlib.reload(core_functional)    # 热更新prompt
-        core_functional = core_functional.get_core_functions()
-        if "PreProcess" in core_functional[additional_fn]: inputs = core_functional[additional_fn]["PreProcess"](inputs)  # 获取预处理函数（如果有的话）
-        inputs = core_functional[additional_fn]["Prefix"] + inputs + core_functional[additional_fn]["Suffix"]
-
-    raw_input = inputs
-    logging.info(f'[raw_input] {raw_input}')
-    chatbot.append((inputs, ""))
-    yield from update_ui(chatbot=chatbot, history=history, msg="等待响应") # 刷新界面
-    
-    payload = generate_azure_payload(inputs, llm_kwargs, history, system_prompt, stream)    
-        
-    history.append(inputs); history.append("")
-
-    retry = 0
-    while True:
-        try:
-            openai.api_type = "azure"            
-            openai.api_version = AZURE_API_VERSION
-            openai.api_base = AZURE_ENDPOINT
-            openai.api_key = AZURE_API_KEY
-            response = openai.ChatCompletion.create(timeout=TIMEOUT_SECONDS, **payload);break
-        except openai.error.AuthenticationError:
-            tb_str = '```\n' + trimmed_format_exc() + '```'
-            chatbot[-1] = [chatbot[-1][0], tb_str]
-            yield from update_ui(chatbot=chatbot, history=history, msg="openai返回错误") # 刷新界面
-            return
-        except:
-            retry += 1
-            traceback.print_exc()
-            if retry > MAX_RETRY: raise TimeoutError
-            if MAX_RETRY!=0: print(f'请求超时，正在重试 ({retry}/{MAX_RETRY}) ……')
-            
-    gpt_replying_buffer = ""    
-    is_head_of_the_stream = True
-    if stream:
-
-        stream_response = response
-
-        while True:
-            try:
-                chunk = next(stream_response)
-                    
-            except StopIteration:                
-                from toolbox import regular_txt_to_markdown; tb_str = '```\n' + trimmed_format_exc() + '```'
-                chatbot[-1] = (chatbot[-1][0], f"[Local Message] 远程返回错误: \n\n{tb_str} \n\n{regular_txt_to_markdown(chunk)}")
-                yield from update_ui(chatbot=chatbot, history=history, msg="远程返回错误:" + chunk) # 刷新界面
-                return            
-            
-            if is_head_of_the_stream and (r'"object":"error"' not in chunk):
-                # 数据流的第一帧不携带content
-                is_head_of_the_stream = False; continue
-            
-            if chunk:
-                #print(chunk)
-                try:                     
-                    if "delta" in chunk["choices"][0]:
-                        if chunk["choices"][0]["finish_reason"] == "stop":
-                            logging.info(f'[response] {gpt_replying_buffer}')
-                            break
-                    status_text = f"finish_reason: {chunk['choices'][0]['finish_reason']}"    
-                    gpt_replying_buffer = gpt_replying_buffer + chunk["choices"][0]["delta"]["content"]                               
-                       
-                    history[-1] = gpt_replying_buffer
-                    chatbot[-1] = (history[-2], history[-1])
-                    yield from update_ui(chatbot=chatbot, history=history, msg=status_text) # 刷新界面
-
-                except Exception as e:
-                    traceback.print_exc()
-                    yield from update_ui(chatbot=chatbot, history=history, msg="Json解析不合常规") # 刷新界面
-                    chunk = get_full_error(chunk, stream_response)
-                    
-                    error_msg = chunk                    
-                    yield from update_ui(chatbot=chatbot, history=history, msg="Json异常" + error_msg) # 刷新界面
-                    return
-
-
-def predict_no_ui_long_connection(inputs, llm_kwargs, history=[], sys_prompt="", observe_window=None, console_slience=False):
-    """
-    发送至AZURE OPENAI API，等待回复，一次性完成，不显示中间过程。但内部用stream的方法避免中途网线被掐。
-    inputs：
-        是本次问询的输入
-    sys_prompt:
-        系统静默prompt
-    llm_kwargs：
-        chatGPT的内部调优参数
-    history：
-        是之前的对话列表
-    observe_window = None：
-        用于负责跨越线程传递已经输出的部分，大部分时候仅仅为了fancy的视觉效果，留空即可。observe_window[0]：观测窗。observe_window[1]：看门狗
-    """
-    watch_dog_patience = 5 # 看门狗的耐心, 设置5秒即可
-    payload = generate_azure_payload(inputs, llm_kwargs, history, system_prompt=sys_prompt, stream=True)
-    retry = 0
-    while True:
-        try:
-            openai.api_type = "azure"            
-            openai.api_version = AZURE_API_VERSION
-            openai.api_base = AZURE_ENDPOINT
-            openai.api_key = AZURE_API_KEY
-            response = openai.ChatCompletion.create(timeout=TIMEOUT_SECONDS, **payload);break
-        except:
-            retry += 1
-            traceback.print_exc()
-            if retry > MAX_RETRY: raise TimeoutError
-            if MAX_RETRY!=0: print(f'请求超时，正在重试 ({retry}/{MAX_RETRY}) ……')
-
-    stream_response =  response
-    result = ''
-    while True:
-        try: chunk = next(stream_response)
-        except StopIteration: 
-            break
-        except:
-            chunk = next(stream_response) # 失败了，重试一次？再失败就没办法了。
-        if len(chunk)==0: continue
-
-        json_data = json.loads(str(chunk))['choices'][0]
-        delta = json_data["delta"]
-        if len(delta) == 0:
-            break
-        if "role" in delta:
-            continue
-        if "content" in delta: 
-            result += delta["content"]
-            if not console_slience: print(delta["content"], end='')
-            if observe_window is not None: 
-                # 观测窗，把已经获取的数据显示出去
-                if len(observe_window) >= 1: observe_window[0] += delta["content"]
-                # 看门狗，如果超过期限没有喂狗，则终止
-                if len(observe_window) >= 2000:  
-                    if (time.time()-observe_window[1]) > watch_dog_patience:
-                        raise RuntimeError("用户取消了程序。")
-        else:
-            raise RuntimeError("意外Json结构："+delta)
-    if json_data['finish_reason'] == 'content_filter':
-        raise RuntimeError("由于提问含不合规内容被Azure过滤。")
-    if json_data['finish_reason'] == 'length':
-        raise ConnectionAbortedError("正常结束，但显示Token不足，导致输出不完整，请削减单次输入的文本量。")
-    return result
-
-
-def generate_azure_payload(inputs, llm_kwargs, history, system_prompt, stream):
-    """
-    整合所有信息，选择LLM模型，生成 azure openai api请求，为发送请求做准备
-    """    
-
-    conversation_cnt = len(history) // 2
-
-    messages = [{"role": "system", "content": system_prompt}]
-    if conversation_cnt:
-        for index in range(0, 2*conversation_cnt, 2):
-            what_i_have_asked = {}
-            what_i_have_asked["role"] = "user"
-            what_i_have_asked["content"] = history[index]
-            what_gpt_answer = {}
-            what_gpt_answer["role"] = "assistant"
-            what_gpt_answer["content"] = history[index+1]
-            if what_i_have_asked["content"] != "":
-                if what_gpt_answer["content"] == "": continue                
-                messages.append(what_i_have_asked)
-                messages.append(what_gpt_answer)
-            else:
-                messages[-1]['content'] = what_gpt_answer['content']
-
-    what_i_ask_now = {}
-    what_i_ask_now["role"] = "user"
-    what_i_ask_now["content"] = inputs
-    messages.append(what_i_ask_now)
-
-    payload = {
-        "model": llm_kwargs['llm_model'],
-        "messages": messages, 
-        "temperature": llm_kwargs['temperature'],  # 1.0,
-        "top_p": llm_kwargs['top_p'],  # 1.0,
-        "n": 1,
-        "stream": stream,
-        "presence_penalty": 0,
-        "frequency_penalty": 0,
-        "engine": AZURE_ENGINE
-    }
-    try:
-        print(f" {llm_kwargs['llm_model']} : {conversation_cnt} : {inputs[:100]} ..........")
-    except:
-        print('输入中可能存在乱码。')
-    return payload
-
-

request_llm/bridge_chatgpt.py

@@ -22,8 +22,8 @@ import importlib
 # config_private.py放自己的秘密如API和代理网址
 # 读取时首先看是否存在私密的config_private配置文件（不受git管控），如果有，则覆盖原config文件
 from toolbox import get_conf, update_ui, is_any_api_key, select_api_key, what_keys, clip_history, trimmed_format_exc
-proxies, API_KEY, TIMEOUT_SECONDS, MAX_RETRY, API_ORG = \
+proxies, TIMEOUT_SECONDS, MAX_RETRY, API_ORG = \
-    get_conf('proxies', 'API_KEY', 'TIMEOUT_SECONDS', 'MAX_RETRY', 'API_ORG')
+    get_conf('proxies', 'TIMEOUT_SECONDS', 'MAX_RETRY', 'API_ORG')
 
 timeout_bot_msg = '[Local Message] Request timeout. Network error. Please check proxy settings in config.py.' + \
                   '网络错误，检查代理服务器是否可用，以及代理设置的格式是否正确，格式须是[协议]://[地址]:[端口]，缺一不可。'
@@ -101,6 +101,8 @@ def predict_no_ui_long_connection(inputs, llm_kwargs, history=[], sys_prompt="",
                     if (time.time()-observe_window[1]) > watch_dog_patience:
                         raise RuntimeError("用户取消了程序。")
         else: raise RuntimeError("意外Json结构："+delta)
+    if json_data['finish_reason'] == 'content_filter':
+        raise RuntimeError("由于提问含不合规内容被Azure过滤。")
     if json_data['finish_reason'] == 'length':
         raise ConnectionAbortedError("正常结束，但显示Token不足，导致输出不完整，请削减单次输入的文本量。")
     return result
@@ -247,6 +249,7 @@ def generate_payload(inputs, llm_kwargs, history, system_prompt, stream):
         "Authorization": f"Bearer {api_key}"
     }
     if API_ORG.startswith('org-'): headers.update({"OpenAI-Organization": API_ORG})
+    if llm_kwargs['llm_model'].startswith('azure-'): headers.update({"api-key": api_key})
 
     conversation_cnt = len(history) // 2
 

toolbox.py

@@ -505,16 +505,24 @@ def on_report_generated(cookies, files, chatbot):
     chatbot.append(['报告如何远程获取？', f'报告已经添加到右侧“文件上传区”（可能处于折叠状态），请查收。{file_links}'])
     return cookies, report_files, chatbot
 
+def load_chat_cookies():
+    API_KEY, LLM_MODEL, AZURE_API_KEY = get_conf('API_KEY', 'LLM_MODEL', 'AZURE_API_KEY')
+    if is_any_api_key(AZURE_API_KEY):
+        if is_any_api_key(API_KEY): API_KEY = API_KEY + ',' + AZURE_API_KEY
+        else: API_KEY = AZURE_API_KEY
+    return {'api_key': API_KEY, 'llm_model': LLM_MODEL}
+
 def is_openai_api_key(key):
     API_MATCH_ORIGINAL = re.match(r"sk-[a-zA-Z0-9]{48}$", key)
+    return bool(API_MATCH_ORIGINAL)
+
+def is_azure_api_key(key):
     API_MATCH_AZURE = re.match(r"[a-zA-Z0-9]{32}$", key)
-    return bool(API_MATCH_ORIGINAL) or bool(API_MATCH_AZURE)
+    return bool(API_MATCH_AZURE)
 
 def is_api2d_key(key):
-    if key.startswith('fk') and len(key) == 41:
+    API_MATCH_API2D = re.match(r"fk[a-zA-Z0-9]{6}-[a-zA-Z0-9]{32}$", key)
-        return True
+    return bool(API_MATCH_API2D)
-    else:
-        return False
 
 def is_any_api_key(key):
     if ',' in key:
@@ -523,10 +531,10 @@ def is_any_api_key(key):
             if is_any_api_key(k): return True
         return False
     else:
-        return is_openai_api_key(key) or is_api2d_key(key)
+        return is_openai_api_key(key) or is_api2d_key(key) or is_azure_api_key(key)
 
 def what_keys(keys):
-    avail_key_list = {'OpenAI Key':0, "API2D Key":0}
+    avail_key_list = {'OpenAI Key':0, "Azure Key":0, "API2D Key":0}
     key_list = keys.split(',')
 
     for k in key_list:
@@ -537,7 +545,11 @@ def what_keys(keys):
         if is_api2d_key(k): 
             avail_key_list['API2D Key'] += 1
 
-    return f"检测到： OpenAI Key {avail_key_list['OpenAI Key']} 个，API2D Key {avail_key_list['API2D Key']} 个"
+    for k in key_list:
+        if is_azure_api_key(k): 
+            avail_key_list['Azure Key'] += 1
+
+    return f"检测到： OpenAI Key {avail_key_list['OpenAI Key']} 个, Azure Key {avail_key_list['Azure Key']} 个, API2D Key {avail_key_list['API2D Key']} 个"
 
 def select_api_key(keys, llm_model):
     import random
@@ -552,8 +564,12 @@ def select_api_key(keys, llm_model):
         for k in key_list:
             if is_api2d_key(k): avail_key_list.append(k)
 
+    if llm_model.startswith('azure-'):
+        for k in key_list:
+            if is_azure_api_key(k): avail_key_list.append(k)
+
     if len(avail_key_list) == 0:
-        raise RuntimeError(f"您提供的api-key不满足要求，不包含任何可用于{llm_model}的api-key。您可能选择了错误的模型或请求源。")
+        raise RuntimeError(f"您提供的api-key不满足要求，不包含任何可用于{llm_model}的api-key。您可能选择了错误的模型或请求源（右下角更换模型菜单中可切换openai,azure和api2d请求源）")
 
     api_key = random.choice(avail_key_list) # 随机负载均衡
     return api_key